Separating Frozen Pages via Learning-Based Recognition with ZNS SSD for Write Amplification Reduction in Database

NAND flash-based solid-state drives (SSDs) are becoming the predominant storage medium for database systems. However, the significant mixing of pages with varying lifespans in B+-tree-based databases exacerbates the problem of huge write amplification (WA) which is caused by the rewrites of live pages during garbage collection (GC) in SSDs, reducing database performance significantly. Frozen pages are dominant in database workloads, which are actively updated for a period and then eventually no longer updated but remain valid. Existing data placement schemes fail to recognize and isolate frozen pages, leading to repeated migrations of them. This paper proposes SepFrozen, a novel machine learning-based mechanism for recognizing and isolating frozen pages to decrease repeated migrations of them, thereby reducing WA and enhancing database performance. SepFrozen can be integrated seamlessly with existing rule-based data placement schemes. We integrate two state-of-the-art schemes with SepFrozen and implement them in LeanStore with ZNS SSDs. Experimental results show that SepFrozen reduces WA by 30.2% and enhances database throughput by 19.5%.